Noise-Clustered Distant Supervision for Relation Extraction: A Nonparametric Bayesian Perspective

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 1808–1813 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Noise-Clustered Distant Supervision for Relation Extraction: A Nonparametric Bayesian Perspective Qing Zhang and Houfeng Wang Key Laboratory of Computational Linguistics (Peking University) Ministry of Education, China {zqicl,wanghf}@pku.edu.cn Abstract For the task of relation extraction, dis- tant supervision is an efficient approach to generate labeled data by aligning knowl- edge base with free texts. The essence of it is a challenging incomplete multi-label classification problem with sparse and noisy features. To address the challenge, this work presents a novel nonparametric Bayesian formulation for the task. Ex- periment results show substantially higher top-precision improvements over the tradi- tional state-of-the-art approaches. 1 Introduction To efficiently generate structured relation informa- tion from free texts, the research on distantly su- pervised Relation Extraction (RE) (Mintz et al., 2009; Riedel et al., 2013; Hoffmann et al., 2011) has been attracting much attention, because it can greatly reduce the manual annotation for training. It essentially based on the assumption that the re- lation between two entities in a Knowledge Base (KB), is also likely hold within a sentence that mentions the two entities in free texts. This as- sumption plays a crucial role in distant supervi- sion, which is quite effective in real applications. However, the assumption of distant alignment can also lead to the noisy training corpus prob- lem (Fan et al., 2014), which is challenging for the task as follows: i) Noisy features. Not all relations existed in a KB keep the same mean- ing of that relation for the corresponding entities in a free text. For example, the second relation mention in Figure 1 does not explicitly describe any relation instance, so features extracted from this sentence can be noisy. Such analogous cases commonly exist in feature extraction. ii) Incom- plete labels. Similar to noisy features, the gener- Figure 1: Aligned Example (Fan et al., 2014): the relation instances related to the entity pair ⟨BarackObama, U.S.⟩ in the KB, and its men- tions in the free text. ated label can be incomplete due to the incomplete knowledge base (Ritter et al., 2013). For exam- ple, the fourth relation mention in Figure 1 should be labeled by the relation Senate-of. However, the corresponding relation instance (Senate-of(Barack Obama, U.S.)) is missing in the knowledge base. Such analogous cases are also common in real ap- plications. iii) Sparse features. Sophisticated features extracted from the mentions can result in a large number of sparse features (Riedel et al., 2013). The generalization ability of feature based prediction models will be badly hurt, when the fea- tures do not match between testing and training. To tackle the problem, we develop a novel dis- tant supervision approach from a nonparametric Bayesian perspective (Blei et al., 2016), along with the previously most effective research line (Petroni et al., 2015) of using matrix completion (Fan et al., 2014) for relation extraction. Our goal is to design a noise-tolerant relation …

Content truncated to 500 words